基于双流自适应特征融合的多模态烟草文档分类  

Multi-Modal Tobacco Document Classification Based on Dual-Stream Adaptive Feature Fusion

作　　者：孙首名 张琦 王喆 苏娜[2] 沈奇 SUN Shouming;ZHANG Qi;WANG Zhe;SU Na;SHEN Qi(Liaoning Tobacco Company Tieling Branch,Tieling Liaoning 112000,China;Northeastern University,Shenyang Liaoning 110000,China)

机构地区：[1]辽宁省烟草公司铁岭市公司,辽宁铁岭112000 [2]东北大学,辽宁沈阳110000

出　　处：《绿洲农业科学与工程》2025年第1期160-163,共4页Oasis Agriculture Science and Engineering

基　　金：国家自然科学基金面上项目一多时延全基因组调控网络的大规模动态概率图建模与分析(62072089)。

摘　　要：针对烟草文档自动化分类的需求,提出一种基于双流自适应特征融合的多模态烟草文档分类网络,名为DSAFFNet。该网络结合烟草文档的文本模态和图像模态,采用DSAFF(Dual-StreamAdaptiveFeatureFusion)模块对不同模态特征的重要性自适应调整权重,实现灵活而精确的多模态融合。试验结果表明,所提网络在烟草文档数据集上的表现优于以往分类方法。To address the need for automated tobacco document classification,we propose DSAFFNet,a multimodal classification network with dual-stream adaptive feature fusion.The network integrates both textual and visual modalities of tobacco documents,utilizing a DSAFF(Dual-StreamAdaptiveFeatureFusion)moduleto adaptively adjusttheimportance weights of different modal features,enabling flexible and accurate multimodal fusion.Experimental results show the proposed network outperforms previous classification methods on the tobacco document dataset.

关 键 词：烟草文档分类 多模态学习 双流网络 

分 类 号：Q953[生物学—动物学]